Biomolecular Simulation: A Computational Microscope for Molecular Biology

7/25/2019 Biomolecular Simulation: A Computational Microscope for Molecular Biology

1/27

Biomolecular Simulation:A Computational Microscopefor Molecular Biology

Ron O. Dror,1 Robert M. Dirks,1 J.P. Grossman,1

Huafeng Xu,1 and David E. Shaw1,2

1D. E. Shaw Research, New York, New York 10036; email: [email protected]@DEShawResearch.com

2Center for Computational Biology and Bioinformatics, Columbia University, New York,New York 10032

Annu. Rev. Biophys. 2012. 41:42952

TheAnnual Review of Biophysicsis online atbiophys.annualreviews.org

This articles doi:10.1146/annurev-biophys-042910-155245

Copyright c2012 by Annual Reviews.All rights reserved

1936-122X/12/0609-0429$20.00

Keywords

molecular dynamics, millisecond timescale, Anton, protein folding, dru

binding

Abstract

Molecular dynamics simulations capture the behavior of biological ma

molecules in full atomic detail, but their computational demands, combwith the challenge of appropriately modeling the relevant physics, have

torically restricted their lengthand accuracy. Dramatic recentimprovemin achievable simulation speed and the underlying physical models

enabled atomic-level simulations on timescales as long as milliseconds

capture key biochemical processes such as protein folding, drug bindmembrane transport, and the conformational changes critical to pro

function. Such simulation may serve as a computational microscopevealing biomolecular mechanisms at spatial and temporal scales that

difficult to observe experimentally. We describe the rapidly evolving of the art for atomic-level biomolecular simulation, illustrate the typ

biological discoveries that can now be made through simulation, and dischallenges motivating continued innovation in this field.

429

Click here for quick links to

Annual Reviews content online,

including:

Other articles in this volume

Top cited articles

Top downloaded articles

Our comprehensive search

FurtherANNUAL

REVIEWS


2/27

Conformationalchange: a transitionbetween twoalternative structuresof a flexiblebiomolecule such as aprotein

Molecular dynamics(MD) simulation: asimulation in whichthe positions andvelocities of atoms arecomputed usingNewtons laws ofmotion

Contents

I N T R O D U C T I O N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 3 0RECENT ADVANCES IN SIMULATION METHODOLOGY................... 433

Accessing Longer Timescales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433Enhanced Sampling and Coarse-Graining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436

Improving Force Field Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436

SIMULATION AS A TOOL FOR MOLECULAR BIOLOGY . . . . . . . . . . . . . . . . . . . . . 437Conformational Changes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437

Membrane Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439Protein Folding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441

Ligand Binding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442FUTURE FRONTIERS OF BIOMOLECULAR SIMULATION. . . . . . . . . . . . . . . . . . 444

Drug Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444Protein Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445

Modeling Nucleic Acids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445Simulation of Complex Biological Structures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446

Enabling Longer and Cheaper Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446

More Accurate Biochemistry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446

INTRODUCTION

Over the past half-century, breakthroughs in structural biology have provided atomic-resolution

models of many of the molecules that are essential to life, including proteins and nucleic acids.

Although static structures determined through crystallography and other techniques are tremen-dously useful, the molecules they represent are, in reality, highly dynamic, and their motions are

often critical to their function (Figure 1). Proteins, for example, undergo a variety of conforma-tional changes that allow them to act as signaling molecules, transporters, catalysts, sensors, and

mechanical effectors. Likewise, they interact dynamically with hormones, drugs, and one anotherStatic structural information might be likened to a photograph of a football game; to understand

more readily how the game is played, we want a video recording.A variety of experimental techniques can provide information about the dynamics of proteins

and other biomolecules, but they are generally limited in their spatial and temporal resolutionand most report ensemble average properties rather than the motion of individual molecules

(Figure 2). An attractive alternative, in principle, is to model atomic-level motions computa-tionally, based on first-principles physics. Although such simulations have been an active area of

research for decades (55), their computational expense, combined with the challenge of develop-

ing appropriate physical models, has placed restrictions on both their length and their accuracyThe past few years have seen great progress in addressing these limitations, making simulations a

much more powerful tool for the study of biomolecular dynamics. This review describes severalimportant recent advances in simulation methodology and offers an overview of what is currently

possible with biomolecular simulation.The quantum mechanical behavior of molecules at a subatomic level is described by the time-

dependent Schrodinger equation, but a direct solution to this equation is in practice computation-ally infeasible for biological macromolecules. The standard method for simulating the motions of

such molecules is a technique known as all-atom molecular dynamics (MD) simulation, in which

430 Dror et al.


3/27

b Ligand bindinga Transport across membrane

c Conformational change d Protein folding

Figure 1

Examples of biomolecular processes that have been examined using molecular dynamics (MD) simulations.(a) Transport of small molecules across the cell membrane. (b) Binding of drugs to their target proteins.(c) Conformational transitions in proteins. (d) Protein folding.

Force field: enerfunction used tocompute the forceacting on atoms (dto interatomicinteractions) durinMD simulation

the positions and velocities of particles representing every atom in the system evolve according

to the laws of classical physics. The forces acting on these particles are computed using a modelknown as a force field, which is typically designed based on a combination of first-principles

physics and parameter fitting to quantum mechanical computations and experimental data.

Although MD simulation does not model the underlying physics exactly, it can provide a suf-ficiently close approximation to capture a wide range of critical biochemical processes. The popu-

larity of such simulations is illustrated by the fact that they account for a majority of the computertime devoted to biomedical research at National Science Foundation supercomputer centers.

Although we briefly touch on approaches that represent molecules at a coarser or finer level ofdetail, or that evolve positions in a nonphysical manner, all-atom MD simulations constitute the

principal focus of this review.Historically, the timescales accessible to MD simulation have been shorter than those on

which most biomolecular events of interest take place, thus limiting the applicability of these

www.annualreviews.org Biomolecular Simulation 431


4/27

Proteintranslation

Activetransport

Actionpotential

Solute permeationProtontransfer

Side chainip

Domainmotion

2 structureformation

Channelgating

Lightabsorption

Electrontransfer

1015 1012 109 106 103 100 103 Static

Time (s)

Length(n

m)

101

100

101

103

104

102

Increasingresolution

Atoms

-sheets

-helices

Proteins

Assemblies

Organelles

Cells

All-atommolecular dynamics

simulations

Electrophysiology

FRET

X-ray

NMR

EM

AFM/optical tweezers

Figure 2

Spatiotemporal resolution of various biophysical techniques. The temporal (abscissa) and spatial (ordinate)resolutions of each technique are indicated by colored boxes. Techniques capable of yielding data on singlemolecules (as opposed to only on ensembles) are in boldface. NMR methods can probe a wide range oftimescales, but they provide limited information on motion at certain intermediate timescales, as indicatedby the lighter shading and dashed lines. The timescales of some fundamental molecular processes, as well ascomposite physiological processes, are indicated below the abscissa. The spatial resolution needed to resolvecertain objects is shown at the right. Adapted from Reference 19. Abbreviations: AFM, atomic forcemicroscopy; EM, electron microscopy; FRET, Forster resonance energy transfer; NMR, nuclear magneticresonance.

simulations. Events such as protein folding, proteindrug binding, and major conformational

changes essential to protein function typically take place on timescales of microseconds to mil-liseconds (Figure 2). MD simulations, by contrast, were until recently generally limitedin practice

to nanosecond timescales. Simulations of even a few microseconds required months on the mostpowerful supercomputers available, and longer simulations had never been performed. Recent

advances in hardware, software, and algorithms have increased the timescales accessible to simula-

tion by several orders of magnitude, enabling the first millisecond-scale simulations and allowingMD to capture many critical biochemical processes for the first time.

The other major factor limiting the applicability of MD has been the accuracy of the force fieldmodels that underlie the simulations. A number of improved force fields have been introduced

432 Dror et al.


5/27

Parallelcomputation: usmultiple cooperatprocessors to perfa computation fastthan would be pos

with a single procMoores law: a trdating back to 196which the logic deon computer chipdoubles approximevery two years

over the past several years, and the longer timescales now accessible to MD simulations have

allowed more extensive validation of these force fields against experimental data.

We begin by summarizing the fundamentals of MD simulation and certain recent methodolog-ical and technological advances that have expanded its applicability. We then review the state of

the art in terms of the types of biological discoveries one can make through simulation, providinga number of recent illustrative examples. Finally, we discuss several classes of important problems

that MD could potentially address in the coming years and the methodological advances that mayhelp solve them.

RECENT ADVANCES IN SIMULATION METHODOLOGY

Although the speed and accuracy of all-atom MD simulations has improved substantially over the

past few years, the basic form of such simulations has endured (1). Each atom in the systemfor

example, a protein and the water surrounding itis represented by a particle (or, in certain cases,multiple particles). The simulation steps through time, alternately computing the forces acting

on each atom and using Newtons laws of motion to update the positions and velocities of all theatoms. Commonly used biomolecular force fields express the total force on an atom as a sum of

three components: (a) bonded forces, which involve interactions between small groups of atomsconnected by one or more covalent bonds; (b) van der Waals forces, which involve interactions

between all pairs of atoms in the system but which fall off quickly with distance and are generallyevaluated only for nearby pairs of atoms; and (c) electrostatic forces, which involve interactions

between all pairs of atoms and fall off slowly with distance. Electrostatic interactions are computedexplicitly between nearby pairs of atoms, whereas long-range electrostaticinteractionsare typically

handled via one of several approximate methods that are more efficient than explicitly computinginteractions between all distant pairs of atoms.

Accessing Longer Timescales

MD simulations are computationally demanding for two reasons. First, the force calculation ateach time step requires substantial computationroughly one billion arithmetic operations for a

system with one hundred thousand atoms. Second, the force calculation must be repeated manytimes. Individual steps are limited to a few femtoseconds by fast atomic vibrations, so simulating a

millisecond of physical time requires nearly one trillion time steps. On a single high-end processorcore, such a simulation would take thousands of years to complete.

To make matters worse, using many computer processors in parallel to accelerate a simulationis challenging. Parallelizing a force calculation across multiple processors requires that those

processors communicate with one another, and the amount of communication required increaseswith the number of processors. Beyond some point, adding more processors actually slows down

the calculation. Furthermore, the force calculations in a single simulation must be performed

sequentially, because the atom positions that serve as input for one force calculation depend onthe results of the previous one. The difficulty of parallelizing an MD simulation implies that the

transistor density improvements predicted by Moores law do not automatically lead to improvedMD performance, because in recent years higher transistor density has translated to more processor

cores per chip, not faster individual processor cores.In spite of these challenges, the recent performance improvements in MD simulations have

far outpaced Moores law. In half a decade, the raw performance of state-of-the-art simulationsincreased by over three orders of magnitude (Figure 3). As recently as 2007, the longest published

all-atom MD simulation of a protein was 2 s; in 2009, the first millisecond-long simulation was



6/27

Performance

(sim

ulatednsday1)

Year

1

10

100

1,000

10,000

2004 2005 2006 2007 2008 2009

Moore's lawtrend

Fastest reported

all-atom MD

simulation

Fastest reportedall-atom MDsimulation

Figure 3

Fastest reported all-atom molecular dynamics (MD) simulations from 2004 to 2009 (blue line). The simulatedsystems ranged from 14,000 to 92,000 atoms, and different simulations were performed using differentparameters, so this data is not intended to be a direct comparison of MD hardware and software systems.Nonetheless, an overall performance trend is evident, substantially exceeding the Moores law growth trendin processing power (black line). The leftmost data point is from a 512-processor simulation using NAMD

(65); the rightmost data point is from a 512-chip simulation on Anton (82). The remaining data points arefrom simulations run using Blue Gene/L (25) and Desmond (9, 14).

published (Table 1). These improvements are attributable to a variety of hardware, software, and

algorithm innovations, which we discuss below.

Parallelization across general-purpose computer chips. Software that parallelizes MD force

calculations across multiple computer processors has existed for two decades (69) but has becomemuch more efficient and scalable in the past several years. IBMs Blue Matter code for its Blue

Gene/L general-purpose supercomputer has been scaled up to 32,768 cores (25). The widelyused MD codes NAMD (65), GROMACS (31), and AMBER (12) have all substantially improved

their parallel performance in recent years. These packages can now deliver performance of over100 ns day1 on commodity computer clusters, with the number of processors required to achieve

this performance scaling roughly linearly with the number of atoms in the system. Desmond, a

software package for commodity clusters developed at D. E. Shaw Research, allowed simulationsnearly an order of magnitude faster than previously possible on the same hardware (9) and has

achieved performance approaching 500 ns day1 (14).These improvements in parallel scalability and efficiency have been possible thanks to a number

of algorithmic innovations, particularly in methods for reducing the communication requirements

Table 1 Longest reported all-atom molecular dynamics simulations from 2006 to 2009

Year Length (s) Protein Platform Reference

2006 2 Rhodopsin Blue Gene/La 54

2007 2 Villin HP-35 GROMACSb 22

2008 10 WW domain NAMDb 27

2009 1,031 BPTI Anton 82

aThis simulation used IBMs Blue Matter software.bThese simulations were performed on a commodity computer cluster with the specified software.

434 Dror et al.


7/27

GPU: graphicsprocessing unit

Special-purposeparallelarchitectures:computerarchitectures desigfor a specific task,often allowing succomputers to comthat task much fasthan general-purpcomputers

Anton: a special-purpose parallelsupercomputerdesigned by D. E.Shaw Research toenable fast MDsimulations

between chips during a simulation. The class of neutral territory methods (9, 10, 25, 81, 86), for

example, substantially reduces the amount of data that must be exchanged between processors in

order to compute range-limited particle interactions.

Graphics processing units. Originally designed specifically to accelerate the rendering of three-

dimensional graphics, graphics processing units (GPUs) have become increasingly popular for

general-purpose scientific computation thanks to their ability to perform large numbers of iden-tical computations in parallel on a single chip. Several MD implementations have been ported to

GPUs (2, 30, 29, 66), and a simulation on one or a few GPUs often rivals the performance of asimulation on a small- to moderate-sized computer cluster. Unfortunately, efficiently parallelizing

across many GPUs is difficult, because communication between GPUs remains slower than com-munication between general-purpose processors; as a result, clusters of GPUs have been unable to

match the performance of large standard clusters. GPUs offer an excellent price-to-performanceratio, however, enabling reasonably fast simulations at a cost substantially lower than that for a

cluster of general-purpose processors.

Special-purpose parallel architectures. By far the greatest recent speedups have been achievedthrough the use of special-purpose chips, in combination with new parallelization algorithms

and software. In particular, a recently developed machine named Anton can perform all-atomMD simulations at up to 20 s day1, approximately two orders of magnitude faster than the

simulation rates generally achieved by any other hardware/software combination (82).Anton is able to achieve this speed because it is designed specifically for MD simulations. The

entire computation is performed on special-purpose chips developed at D. E. Shaw Research,

which are directly connected to one another in a custom network (Figure 4). Each Anton chipcontains an array of arithmetic units hardwired for computing particle interactions, enabling a

single Anton chip to compute interactions hundreds of times faster than a commodity proces-sor core (82). Each chip also contains a dozen programmable processor cores, customized for

MD, which accelerate the remainder of the computation. In addition, specialized data-movement

a b

Figure 4

Anton, a special-purpose computer for molecular dynamics designed by D. E. Shaw Research, has performed all-atom proteinsimulations over one hundred times longer than any published previously. ( a) A single Anton chip. (b) The first Anton machine,comprising 512 Anton chips connected through a specially designed network.



8/27

Enhanced sampling:algorithms devised tospeed up theexploration ofmolecularconformations by

altering the physics ofthe system

operations support fast communication between small on-chip memories, eliminating the memory

cache hierarchy that typically consumes the majority of the area on commodity chips.

Several algorithmic advances also contribute to Antons performance. A specific neutral terri-tory method (81) was designed for Anton and is directly implemented within its specialized parti-

cle interaction hardware. Anton computes long-range electrostatic forces using the Gaussian splitEwald method (79) rather than the more commonly used particle mesh Ewald method, allowing

a significant portion of the long-range electrostatics computation to be performed by the samespecialized hardware that handles particle interactions. Finally, the communication patterns in

Antons MD software, which differsignificantly from those in other parallel MD software packagesare designed to take advantage of Antons specialized low-latency mechanisms for communication

between and within chips (18).Several previous projects, including FASTRUN (24), MD Engine (92), and MDGRAPE (90),

have built special-purpose hardware to accelerate the most computationally expensive elements

of an MD simulation. Although such hardware reduces the effective cost of simulating a givenperiod of biological time, the speedup achieved through parallelization across many such chips is

limited by the remainder of the computation as well as the communication required, precludingindividual simulations on multi-microsecond timescales.

Anton has enabled all-atom MD simulations of proteins of more than a millisecond in lengthover 100 times longer than any such simulation reported using other hardware. With Anton it

becomes possible, forthe first time, to directly observe in simulation various importantbiochemicaprocesses that occur on timescales greater than a few microseconds.

Enhanced Sampling and Coarse-Graining

A variety of methods can be used in combination with MD simulation to investigate events that

occur on timescales that remain inaccessible to direct all-atom MD simulation. Zuckerman (106)

provided a thorough review of such enhanced sampling methods in Volume 40 of the AnnualReview of Biophysics, in which he concluded that No other method can routinely and reliably

outperform MD by a significant amount. These methods prove essential in certain situationshowever, as illustrated by several of the case studies in the following section.

One can also sidestep communication bottlenecks by performing many short simulations inparallel. A particularly impressive example is the Folding@Home project (5), which has obtained

significant scientific results by using over 400,000 personal computers (PCs) to simulate manyseparate molecular trajectories, each limited to the timescale accessible on a single PC. The

solution of many important problems, however, is greatly facilitated by the availability of longindividual trajectories.

Finally, one can often substantially accelerate simulations at the cost of reduced accuracyby employing simplified system representations, such as coarse-grained models, in which each

simulated particle represents several atoms, or implicit solvent models in which water atoms are

replaced by a continuum representation. Models of both types have seen substantial developmentin recent years (13, 58).

Improving Force Field Accuracy

Long-timescale simulations place more stringent demands on force fields; a force field that proves

sufficient for short-timescale simulations may not be sufficient at longer timescales. Fortunately,the past several years have seen substantial improvements in force fields for biomolecular simula-

tion. Historically, force field parameters were determined using quantum mechanical calculations

436 Dror et al.


9/27

G-protein-couplreceptors (GPCRa family oftransmembraneproteins that transsignals into cells a

represent the largclass of drug targe

and condensed phase experimental data for small molecular fragments. Recently, force field de-

velopment has come to increasingly rely on experimental data for proteins and other biological

macromolecules, as improvements in both simulation speed and experimental methods have ledto an overlap in the timescales accessible through the two approaches.

Although the functional forms of the most widely used force fields have remained largelyunchanged, their parameters have recently undergone a number of adjustments. The Amber

force field, for example, incorporated changes to parameters associated with torsional angles ofthe protein backbone, first to improve fits to quantum calculations (32) and then to achieve

better agreement between secondary structure preferences observed in long MD simulations ofpolypeptides and corresponding NMR measurements (7). Amber protein side chain torsions were

also adjusted to better match both quantum calculations and NMR data (51). Adjustments tobackbone and side chain torsions were also incorporated into the CHARMM force field (53, 67),

as were modifications to the charge distributions of ionizable amino acid residues (67). Recent

studies have also resulted in improved parameters for lipids in CHARMM (42) and for smalldrug-like molecules in the CHARMM, Amber, and OPLS-AA force fields (4, 96, 98).

A recent study exploited long-timescale MD simulations on Anton to evaluate a number ofprotein force fields through a systematic comparison with experimental data (49). Criteria in-

cluded the ability of each force field to fold small proteins to their native structures, to predict thesecondary structure propensities of polypeptides, and to reproduce NMR data reporting on the

structure and dynamics of folded proteins. The results indicated that the force fields examinedhave consistently improved overthe past decade, andthat themost recentversions providean accu-

rate description of many structural and dynamic properties of proteins. The study also highlightedcertain shortcomings: None of the force fields, for example, were able to accurately capture the

temperature dependency of the secondary structure propensities. It is an open question whether

the ongoing parameterization of existing functional forms will be sufficient to further improveforce fields. Substantial efforts are under way to develop force fields with more sophisticated

functional forms, including polarizable force fields (36, 70), which capture the redistribution ofelectrons around each atom in response to changes in environment.

SIMULATION AS A TOOL FOR MOLECULAR BIOLOGYIn this section, we illustrate the utility of modern MD simulation as a biological research tool

through a number of recent case studies, many of which are drawn from our own work, involvingconformational changes in proteins, transport across membranes, protein folding, and the binding

of ligands to proteins (Figure 1).

Conformational Changes

Under physiological conditions, proteins and other biomacromolecules constantly move from

one structural state to another, and their function and regulation depend on these conformationalchanges. MD simulations are often used to identify novel conformations, to capture the transi-

tional pathways between conformations, to determine equilibrium distributions among differentconformations, and to characterize changes in conformational distribution as a result of mutation

or ligand binding. We provide several examples involving G-protein-coupled receptors (GPCRs)and kinases.

GPCRs represent the largest class of drug targets: One-third of all marketed drugs act bybinding to a GPCR and either triggering or preventing receptor activation, which involves

a transition from an inactive receptor conformation to an active conformation that causes



10/27

2-adrenergicreceptor (2AR): anarchetypal GPCR anda target of betablockers and betaagonists

G-protein-mediated signaling. The past few years have witnessed the determination of the first

several crystal structures of ligand-activated GPCRs, beginning with the 2-adrenergic receptor

(2AR). MD has addressed several key questions raised by these structures about the conforma-tions of inactive states and the mechanism of receptor activation (16, 17, 52, 62, 72, 73, 95).

An MD simulation study by Dror et al. (16) identified a previously unobserved inactive con-formation of2AR, resolving an apparent contradiction between experimental results: A network

of salt bridges known as the ionic lock, suggested by biochemical experiments to stabilize theinactive state of2AR and other GPCRs, was disrupted in the inactive state crystal structures

(46). In microsecond-timescale simulations of inactive2AR, the receptor transitions repeatedlybetween two conformational states, one with the lock broken and one with it formed. In simu-

lations of wild-type 2AR, the lock-formed conformation predominates, in agreement with thebiochemical data, but in simulations of the modified protein used for crystallographic structure

determination, the lock-broken conformation predominates, explaining the crystallographic re-

sult. Both lock-broken and lock-formed conformations were subsequently observed in a crystalstructure of a closely related receptor (60), lending support to these computational predictions.

More recent simulations on Anton (73) captured spontaneous transitions of2AR from an activeto an inactive conformation, addressing a puzzle posed by two recent crystallographic structures of

2AR bound to two differentagonists (ligands that cause activation). One of these structures, whichalso has a G-protein-mimetic nanobody bound to its intracellular surface, appears to represent an

active conformation (71). The other, which is bound to an agonist but lacks the nanobody, is almostidentical to the previously solved inactive structure (73). Is this surprising structural difference due

to differences between the agonists or the crystallized receptor constructs, or might it be due to thepresence or absence of the G-protein-mimetic partner? In multi-microsecond simulations of2AR

initiated from the nanobody-bound active structure, but with the nanobody removed, the agonist-bound receptor spontaneously transitioned to a conformation that closely matched the inactive

crystal structure. Taken together, these simulations and the crystal structures suggest that, even

with an agonist bound, the majority of the 2AR population remains in an inactive-like conforma-tion until a G-protein or G-protein-mimetic nanobody binds, stabilizing the active conformation

These simulationswhich represent the first in which a GPCR transitions spontaneouslybetween crystallographic conformations representing functionally distinct statesalso served to

characterize the atomic-level activation mechanism of2AR (17). They showed that the extra-cellular drug-binding site is connected to the intracellular G-protein-binding site via a loosely

coupled allosteric network, comprising three regions that can switch individually between dis-tinct conformations. The simulations revealed a key intermediate conformation on the activation

pathway and suggestedsomewhat counterintuitivelythat the first structural changes duringactivation often take place on the intracellular side of the receptor, far from the drug-binding site.

These results may provide a foundation for the design of drugs that control receptor signalingmore precisely by stabilizing specific receptor conformations.

MD simulations have also shed light on the function and regulation of kinases, a class of

enzymes that are actively pursued as therapeutic targets for cancer and autoimmune diseases. Theactivity of a typical kinase is regulated by changes to its preference for the active and various

inactive conformations. Mutations to kinases often favor the wrong conformations, leading toaberrant signaling and consequently disease. A number of studies have used MD simulations to

characterize conformational changes in kinases (6, 23, 80, 100), yielding predictions that were inagreement with subsequent experimental measurements (80) and insights that led to the design

of new experimental methods (76).Faraldo-G omez & Roux (23) used MD simulations to characterize the regulation of Src family

tyrosine kinases, which depends on an inactivation process in which the auxiliary domains (known

438 Dror et al.


11/27

Membranetransport:the movement ofmolecules across amembrane, usuallyfacilitated by a

transmembraneprotein

as SH2 and SH3) of a kinase assemble onto its catalytic domain, preventing catalysis. What

makes such assembly robust and fast, so that kinases can be reliably and quickly turned off? To

answer these questions, the authors used an enhanced sampling technique known as umbrellasampling (44) to characterize the relative free energies of various conformations connecting the

disassembled and assembled states. The simulations indicated that the SH2SH3 construct hasan intrinsic propensity to adopt conformations primed for association with the catalytic domain,

thus favoring and accelerating formation of the assembled (inhibitory) state. Their results alsosuggested that the SH2SH3 connector is more than a passive link between the domains; rather,

it is responsible for their propensity toward the assembly-ready conformation, explaining theexperimental observation that mutations in the connector region increase the constitutive activity

of the kinase.

Membrane Transport

Transport of various substrates across the cell membrane is vital both to maintaining a cells con-

stitution and to transmitting biochemical signals. The transport efficiency and substrate selectivityof carrier proteins often depend critically on the detailed spatial configuration of the atoms along

the transport pathway as well as their subtle movement during the transport process. MD sim-ulation, with its unique ability to simulate and record the movement of individual atoms at very

fine temporal and spatial resolutions, lends itself naturally to the study of transport processes. Inthe past decade, MD simulations have been applied to investigate a number of transporters and

channels, including aquaporins (35, 91), ion channels (8, 34, 63), and active transporters (3, 21).These studies have shed light on many mechanistic questions: How do the channels achieve a fast

rate of substrate permeation? How do the transporters affect selectivity for their substrates? Howis transport regulated in response to various stimuli?

Potassium channels, which allow potassium ions to move passively through the cell membrane,

are essential for the transmission of nerve impulses and represent an important target for thetreatment of diseases ranging from Alzheimers to diabetes. A longstanding puzzle about these

channels is why they let potassium ions, but not smaller sodium ions, pass through. Crystal struc-tures suggest that the narrowest region of a potassium channel, known as the selectivity filter,

has a geometry that snugly fits potassium ions, but the difference between the radii of potassiumand sodium ions (0.38 A) is smaller than the thermal fluctuations of the atomic positions in the

selectivity filter (0.75 A). Noskov et al. (63) and Bostick & Brooks (8) addressed this question byusing MD simulations to examine several hypothetical variants of the real selectivity filter. In both

studies, the authors computed the difference in the binding free energies of sodium and potassiumions to the selectivity filter and explored how this difference varied when they artificially adjusted

the physical properties of the filter. Both studies suggested that selectivity was a robust featureof the filter that did not depend on its precise geometry. Instead, selectivity was a consequence of

the dipole moment of the carbonyls coordinating the ions in the selectivity filter, the coordination

number, and the thermal fluctuations in the filter.Recent advances in simulation speed have allowed the first direct, atomic-resolution ob-

servations of ion permeation and pore domain closure in a voltage-gated potassium channel(Figure 5). Using unbiased microsecond-timescale MD simulations at various transmembrane

voltages, Jensen et al. (34) followed the permeation of hundreds of potassium ions through thechannel. The authors identified the transitions between microscopic states that underlie the per-

meation of an individual ion, thereby supplying atomistic detail of the long-postulated knock-onconduction mechanism, in which translocation of two selectivity-filter-bound ions is driven by a

third, incoming ion.



12/27

Closed andfully dewetted

Closed andpartially dewetted

0

10

10

5

5

Ionpositio

nalongporeaxis()

0.1 0.2 0.3 0.4

Time (s)

H

2

OH2O

K

K+

Time

a Ion transport at positive transmembrane potential

Open andwetted

b Negative transmembrane potential induces channel closure

Figure 5

Simulation of ion permeation and gating in a potassium channel. ( a) Potassium ions permeated outward (inthe figure, upward) through the selectivity filter when the transmembrane potential was positive. Individualions paused at well-defined sites within the filter, as shown by the representative traces in green. ( b) Whenthe transmembrane voltage was reversed, the hydrophobic cavity dehydrated, causing it to collapse and thusclose the channel to conduction. Figure adapted from Reference 34.

Moreover, Jensen et al. observed channel closuregating of the potassium channel poredomainat negative voltages (Figure 5). Closure took place by means of a previously hypoth-

esized, but unobserved (for ion channels) mechanism, called hydrophobic gating, in which thehydrophobic cavity adjacent to the selectivity filter dehydrated, causing the open pore domain to

collapse into a closed conformation. This mechanism provides a molecular explanation for theexperimental observation that the channel conductance is sensitive to the osmotic pressure. In

particular, the change in volume upon channel closure has been measured experimentally, and

it corresponds to the volume of 4050 water molecules, closely matching the number of watermolecules expelled from the pore cavity upon channel closure in the simulations (105).

440 Dror et al.


13/27

Folding pathwaya sequence ofintermediatestructures visited bprotein as it transifrom a disordered

to its native state

MD simulations have also been used to deduce the mechanism of the sodium proton antiporter,

NhaA (3), a transporter that moves sodium ions and protons in opposite directions across the cell

membrane. Arkin et al. (3) performed a series of simulations in which they systematically variedthe initial position of the sodium ion, as well as the protonation states of two aspartate residues

Asp163 and Asp164critical for antiporting function. These simulations suggested that Asp164serves as the binding site of sodium ion, with its protonation state determining whether the ion will

remain bound or be released into the membrane, and that Aps163 acts as an accessibility controlsite, determining whether the ion will be released to the inside or outside of the cell. Although

the simulations (100 ns each) were much shorter than the complete antiporting cycle (1 ms),they allowed formulation of a complete transport mechanism, which was substantiated through

experiments on NhaA mutants.

Protein Folding

Protein folding actually represents two challenges: Given only a proteins amino acid sequence,

(a) determine the native structure of the protein, and (b) elucidate the pathways by which it folds tothat structure. MD can potentially address both challenges (87), but it is particularly well suited for

revealing folding pathways. Many computational (43) and experimental methods directly predict

or determine protein structure, but few techniques allow direct observation of the dynamics of afolding event in atomic detail. Given an accurate force field and sufficient simulation time, MDcan produce atomic-level trajectories of spontaneous folding events (22, 28, 37, 47, 8385, 97), as

well as unfolding events (93). Such a microscopic view can shed light not only on the structure andstability of the folded state, but also on the heterogeneity of folding pathways, the rate-limiting

steps on these pathways, the nature of misfolded states, and other complex features of the proteinfolding process.

Improvements in both simulation speed and force field accuracy recently enabled Lindorff-

Larsen et al. (50) to simulate repeated folding and unfolding events for a structurally diverse setof 12 small, fast-folding proteins, using a single force field. All 12 proteins folded to structures

closely resembling those determined experimentally (Figure 6). The ability of simulations to

identify the native structures is itself noteworthy, suggesting that MD may eventually serve as aviable method for predicting or refining the structures of arbitrary proteins. The most immediateutility of these simulations, however, is in allowing direct observation of the protein folding

process.Comparing the behavior of these 12 proteins suggested unifying principles for protein folding,

at least for small, fast-folding proteins, and allowed the authors to address several long-standingquestions regarding the mechanisms of protein folding (88). Most of the proteins studied, for

example, consistently fold along a single dominant route, with local structures forming in an order

that largely corresponds to the stability of those structures in the unfolded ensemble. In addition,a few long-range contacts typically form early in the folding process and help establish a nucleus

to guide formation of the rest of the structure.

MD can also help guide wet-lab protein folding experiments (45). Piana et al. (68), for exam-ple, used insights gained from long simulations of a WW domain to suggest a triple mutationthat should reduce the main energy barrier on the folding pathway and thus accelerate fold-

ing. Temperature-jump experiments confirmed this prediction, establishing this mutant as thefastest folding-sheet protein knowna conclusion made more noteworthy because substantial

effort had previously been dedicated to maximizing the folding rate of this WW domain throughmutagenesis (61).



14/27

Chignolin BBATrp-cage Villin

WW domain NTL9 BBL Protein B

Homeodomain Protein G 3D repressor

Figure 6In simulations with a single force field, 12 structurally diverse proteins fold spontaneously to a structure(blue) closely resembling that determined experimentally (red). The simulation snapshots were chosenautomatically based on a clustering analysis that did not exploit knowledge of the experimental structure.The total simulation time per protein ranged from 104 to 2,936 s, allowing observation of at least 10folding and 10 unfolding events for each protein. Figure adapted from Reference 50.

Ligand Binding

Interactions between proteins and small-molecule ligands play a key role in intercellular signaling

and, when the ligands are drugs, in the treatment of disease. Ligands typically affect proteinfunction by directly blocking the active site of a protein or by causing the protein to adopt a

functionally altered conformational state.Thanks to recent advances in accessible timescales, it is now possible to perform MD simula-

tions in which ligands bind spontaneously to proteins without any prior knowledge of the bindingsite (20, 33, 78). In work by Shan et al. (78) on inhibitors binding to Src kinase, and by Dror et al.

(20) on beta blockers and a beta agonist binding to two GPCRs, simulated drug molecules diffusedextensively about the protein before discovering their binding site and binding in a location and

conformation that match crystallographic observations almost exactly (Figure 7). These results

442 Dror et al.


15/27

(S)-dihydroalprenolol

NH2+

HO

O

0.51 s 0.68 s 0.80 s 1.88 s 3.68 s

5

10

15

20

25

30

Ligandrmsd()

0 1 2 3 4 Time (s)

b

Extracellular spaceExtracellular space

a5Time (s)0

3

45

1 2

Extracellular

vestibule

Extracellularvestibule

54

1 2 3 4

2

3

1

Figure 7

Beta blockers bind spontaneously to the 2-adrenergic receptor (2AR) in molecular dynamics simulations, achieving thecrystallographic pose and revealing several metastable intermediate states on the binding pathway. ( a, top left) Pins indicate success

positions of a dihydroalprenolol molecule as it binds to 2AR. The ligand moves from bulk solvent (pose), into the extracellulavestibule (poses and), and finally into the binding pocket (poses and). (a, bottom) These five poses are shown in purple,the crystallographic pose in gray. (a, top right) The path taken by the ligand as it diffuses about the receptor and then binds.(b) Root mean square deviation (rmsd) of the ligand in simulation from its position in the alprenolol2AR crystal structure. Figuradapted from Reference 18.



16/27

raise the possibility of using simulation to identify novel binding sites. Indeed, both Shan et al. and

Dror et al. discovered alternative binding sites, suggesting possibilities for the design of allosteric

drugs with improved selectivity among kinases or GPCRs.Such simulations also allow atomic-level characterization of the binding pathways andenergetic

barriers that determine binding kinetics. Dror et al. (20) found that beta blockers visit a sequenceof metastable conformations en route to the binding pocket of the 2AR (Figure 7). Surprisingly

they found that the largest energetic barrier on the binding pathway often occurs much earlier thanreceptor geometry would suggest, and appears to involve substantial dehydration that occurs as

the drug associates with a particular region on the receptor surface. Shan et al. (78) also identifiedmetastable conformations on the binding pathway, as did Buch et al. (11) in a study of an inhibitor

binding to trypsin. These studies are computationally intensive: Shan et al., Dror et al., and Buchet al. performed multiple simulations totaling over 150 s, 400s, and 50s, respectively.

A common application of proteinligand simulations is to compute the binding affinity of a

ligand, often a drug candidate,to a known binding site. Unbiased MD simulations of ligandbindingare usually ill suited for this purpose, as precise estimation of ligand affinity would typically require

seconds to hours of simulated time in order to observe sufficiently many binding and unbindingevents. Fortunately, binding affinity calculations can be performed much more efficiently using

methods such as free energy perturbation (107) or thermodynamic integration (40), which involveusing a family of modified force fields to bias a series of simulations in ways that accelerate

the forming and breaking of protein-ligand interactions. These biasing forces can be physicallyintuitive, such as forcibly pulling a ligand into or out of a known binding pocket (99), or more

abstract, such as gradually turning off all interactions between a ligand and its surroundings (38). Ifthe artificial energy functions are properly constructed, unbiased binding affinities can be efficiently

and quantitatively derived from the biased simulations.

One compelling example of simulation-based binding affinity calculations is recent work onHIV reverse transcriptase. Starting with a weakly binding ligand that displayed no activity as a

reverse transcriptase inhibitor, Zeevaart et al. (103) used Monte Carlo simulations (closely relatedto MD) to calculate the relative binding energies of a family of closely related molecules. By

selecting variants predicted to bind more tightly, they discovered several molecules that provedexperimentally active in protecting human T-cells from HIV infection.

FUTURE FRONTIERS OF BIOMOLECULAR SIMULATION

As MD simulations become faster and more accurate, they will almost certainly find applications

beyond the categories we have discussed. In this section, we speculate on other areas on whichbiomolecular simulation may make an impact in the coming years, and highlight some of the

methodological problems whose solutions may help make these possibilities reality.

Drug Design

A major goal of structural biology in general, and biomolecular simulation in particular, has long

been to assist in the design of therapeutics. Simulations are already sometimes utilized as part ofthe drug development process. Simulation-based binding affinity calculations, for example, guided

the design of HIV reverse transcriptase inhibitors mentioned above (103), as well as the subsequentdesign of inhibitors that maintain high potency in the presence of a drug-resistance mutation (37)

The use of MD in mainstream drug discovery efforts, however, remains limited.In the future, simulation may offer a number of opportunities for improving the drug discovery

process. Simulation-based methods can compute ligandprotein affinities more accurately than

444 Dror et al.


17/27

standard docking methods, aiding in the identification of lead compounds through virtual screen-

ing of drug candidates or through a fragment-based approach. Accurate evaluation of binding

affinities may prove even more useful in the subsequent process of lead optimization, or in avoid-ing toxicity by ensuring that drug candidates do not bind to known antitargets. MD also has the

potential to discover novel binding sites, including pockets that are not present in existing crystalstructures (75, 78). In addition, simulations may allow refinement of low-resolution structural

models for proteins, thus enabling structure-based drug design.MD also allows the examination of interactions between known drugs and genetic variants

of protein targets. If a disease becomes resistant to a drug, simulations of the mutated targetsmay elucidate the mechanism of resistance and facilitate modifications that restore drug efficacy

(37). Simulations might even aid in the design of drugs or drug cocktails tailored to the geneticmakeupand thus the unique protein variantsof a particular individual.

Perhaps more importantly, the insights MD can provide into the functional mechanisms of

proteins involved in disease pathways may facilitate the identification of appropriate targets andthe design of drugs that target those proteins. Many drugs may prove more effective if they bind

preferentially to a specific conformation of their target protein. Such conformational selectivitycould allow finer control of cellular signaling by stabilizing a particular conformation of a re-

ceptor, or reduce side effects by favoring binding to a protein when it is in a particular state ofactivity.

Protein Design

By facilitating optimization of properties such as structure, ligand-binding affinity, or enzymaticactivity, MD may play a role in the design of proteins for use as biosensors, industrial catalysts,

or therapeutic antibodies, among other potential applications. MD has already been used to rankcandidate amino acid sequences on the basis of calculated properties such as binding affinity (41,

59, 101). It may be used in the future not only to test whether a protein binds a ligand, forms a

desired interface with another protein, or folds correctly, but to guide the design process in orderto achieve such properties. One might even imagine a simulation during which a protein gradually

evolves, favoring mutations that improve some measure of its fitness.

Modeling Nucleic Acids

Although nucleic acids are often viewed as static structures that primarily carry sequence infor-

mation, experimental investigations of both RNAs and DNAs, such as those found in ribosomes,riboswitches, Holliday junctions, and nucleosomes, reveal diverse structures, functions, and dy-

namics. As with proteins, accurate, long-timescale simulations of nucleic acids should enrich ourunderstanding of important biological mechanisms and may help reveal promising drug targets

and binding sites.

To date, MD simulations of proteins have vastly outnumbered those of nucleic acids

(56). Reasons for this disparity include both the smaller number of available nucleic acidstructures and the relative immaturity of nucleic acid force fields. The Protein Data Bank(http://www.pdb.org/)currently contains over 75,000 solved structures, whereas the Nucleic

Acid Database(http://ndbserver.rutgers.edu/)contains just over 5,000. Force fields for nucleicacids are still undergoing extensive refinement (15, 64, 102, 104), and it is unclear whether the

high charge density of nucleic acids and commonly associated divalent cations can be accuratelysimulated using simple point charges; polarizability may be essential.

http://www.pdb.org/http://www.pdb.org/http://ndbserver.rutgers.edu/http://ndbserver.rutgers.edu/http://ndbserver.rutgers.edu/http://ndbserver.rutgers.edu/http://www.pdb.org/


18/27

Simulation of Complex Biological Structures

The increasing availability of high-resolution structures for biological units of greater size andcomplexitymotor proteins, multiprotein complexes, ribosomes, whole organelles, and even sim-

ple organismsoffers the opportunity to investigate their dynamics through simulation. All-atomMD simulations have already been employed to model a few such units. Using simulations of more

than two million atoms, Sanbonmatsu et al. (74) investigated howa ribosome selectstransfer RNAs(tRNAs) with high efficiency and accuracy and found that the flexibility of tRNAs was crucial to

maintaining the complementary codonanticodon geometry. Freddolino et al. (26) conductedMD simulations of the complete satellite tobacco mosaic virus, including a capsid consisting of

60 copies of a single protein and a 1,058-base RNA genome. These simulations, with more than a

million atoms, suggested that the presence of the RNA is necessary for the assembly and stabilityof the capsid.

Although such million-atom simulations are impressive, cellular organelles, let alone entirecells, are dramatically larger; a mitochondrion, for instance, is about half a micron in diameter and

comprises over ten billion atoms. Further, the functional timescales of large protein complexesand organelles tend to be substantially longer than those of individual proteins, often extending

to seconds or more. Such spatial and temporal scales are well beyond those of even the mostadvanced MD simulations. Fortunately, complex biological structures usually have a hierarchical

and modular organization; it may thus be especially productive to develop multiscale models thatuse the most appropriate abstraction and representation for each temporal and spatial scale. The

challenge lies in integrating all-atom MD simulations seamlessly into such multiscale models.

Enabling Longer and Cheaper Simulations

Although the millisecond-scale simulations recently performed on Anton represent a significantmilestone, many biochemical events take place on timescales well above a millisecond and involve

systems larger than those currently being simulated. The efficiency of a parallel MD simulationon a given hardware platform is determined roughly by the ratio of atoms to processors, so to first

approximation one can roughly maintain simulation speed for larger chemical systems by using

additional processors. In practice, however, cost often becomes a limiting factor. Accessing longertimescales is even more difficult: Because of interprocessor communication limits, one cannotincrease the speed of todays fastest simulations by simply adding more processors. Innovation

in some combination of algorithms, computer architecture, and enhanced sampling methods isthus necessary to reach longer timescales and to reduce the cost of large simulations. Such work

is currently under way, both in our group and elsewhere.

More Accurate Biochemistry

Despite recent improvements in force fields, MD simulations still face a number of sources oferror. First, force fields remain imperfect. Second, most contemporary MD simulations do not

fully capture the detailed molecular composition of biological systems, in which many differenttypes of molecules are present, both in the aqueous phase and in lipid bilayers. In the past, force

field inaccuracies often overshadowed the effects of unrealistic composition, but as force fieldsimprove, accurately modeling molecular composition will become more important.

Third, classical MD simulations treat covalent bonds as unchanging. Chemical reactions, inwhich covalent bonds break or form, are typically simulated using techniques such as quantum

mechanics/molecular mechanics simulations, in which the bulk of the system is simulated as

446 Dror et al.


19/27

in classical MD, but a small part is evaluated using more computationally intensive quantum

mechanical approaches (39, 77). Several methods are under development to handle chemical

reactions directly within an MD framework. Some of these concentrate on capturing changes tothe protonation states of ionizable amino acid residues (57, 89). Reactive force fields, which allow

covalent bonds between arbitrary pairs of atoms to break or form, have thus far been limited tosimple inorganic and organic molecules (94) but may eventually capture more general enzymatic

reactions.A promising avenue to improve the accuracy of MD simulations is to incorporate experimen-

tal data directly into the simulations. NMR data, for example, has been used to restrain MDsimulations, biasing the protein conformations toward those compatible with the experimental

measurements (48). A general framework allowing incorporation of biophysical, biochemical, andeven evolutionary data into MD simulations may prove useful both in interpreting experimental

data and in making simulation-based predictions.

SUMMARY POINTS

1. MD simulation can serve as a computational microscope, revealing the workings of

biomolecular systems at a spatial and temporal resolution that is often difficult to accessexperimentally.

2. Until recently, even the longest atomic-level MD simulations fell short of the microsec-ond and millisecond timescales on which biochemical events such as protein folding,

proteindrug interactions, and major conformational changes typically take place. Thespeed of the fastest MD simulations has increased 1,000-fold over a period of several

years, however, due to the development of specialized hardware and better paralleliza-

tion algorithms. All-atom simulations of proteins can now reach timescales in excess ofa millisecond.

3. These developments, combined with the improvements to the force field models that

underlie MD simulations, have allowed MD to capture in atomistic detail processes such

as the conformational transitions essential to protein function, the folding of proteins to

their native structures, the transport of small molecules across cell membranes, and thebinding of drugs to their targets.

FUTURE ISSUES

1. Many biochemical events still take place on long timescales that areinaccessibleto atomic-

level MD simulations, or on large spatial scales that make atomic-level simulation inor-dinately expensive. Further improvements in algorithms and computer architectures are

needed to make simulations faster and more cost-effective.

2. Multiscale models and enhanced sampling methods will likely also play an essential role

in capturing events at larger temporal and spatial scales.

3. Force fields require further improvement and validation, particularly for the modeling

of nucleic acids, certain ions, and some types of ligands.

4. Classical MD does not capture breaking and formation of covalent bonds, but it may be

possible to handle such reactive chemistry within a generalized MD framework.



20/27

5. Application of MD simulation to the design of drugs and proteins remains fertile ground

for future research.

DISCLOSURE STATEMENT

David E. Shaw is the beneficial owner of D. E. Shaw Research and serves as its Chief Scientist.

ACKNOWLEDGMENTS

We thank Anton Arkhipov, David Borhani, Michael Eastwood, Morten Jensen, John KlepeisKresten Lindorff-Larsen, Venkatesh Mysore, Albert Pan, Stefano Piana, and Yibing Shan for

stimulatingdiscussions, helpful comments, and assistance with figures, and Mollie Kirkfor editoriaassistance.

LITERATURE CITED

1. Adcock SA, McCammon JA. 2006 . Molecular dynamics: survey of methods for simulating the activity

of proteins.Chem. Rev.106:1589615

2. Anderson JA, Lorenz CD, Travesset A. 2008. General purpose molecular dynamics simulations fully

implemented on graphics processing units.J. Comp. Phys.227:534259

3. Arkin IT, Xu H, Jensen M, Arbely E, Bennett ER, et al. 2007. Mechanism of Na+/H+ antiporting

Science317:799803

4. Banks JL, Beard HS, Cao Y, Cho AE, Damm W, et al. 2005. Integrated modeling program, applied

chemical theory (IMPACT).J. Comp. Chem.26:175280

5. Folding@home uses

spare cycles on over

400,000 processors

distributed across the

internet to perform MD

simulations.

5. Beberg AL, Ensign D, Jayachandran G, Khaliq S, Pande VS. 2009. Folding@home: lessons from

eight years of volunteer distributed computing. Proc. 2009 IEEE Intl. Symp. Parallel Distributed

Processing(IPDPS 09). Washington, DC: IEEE Comp. Soc.

6. Berteotti A, Cavalli A, Branduardi D, Gervasio FL, Recanatini M, Parrinello M. 2009. Protein confor-

mational transitions: the closure mechanism of a kinase explored by atomistic simulations.J. Am. ChemSoc.131:24450

7. Best RB,Hummer G. 2009. Optimizedmolecular dynamics force fields applied to thehelix-coil transition

of polypeptides.J. Phys. Chem. B113:900415

8. Bostick DL,Brooks CL III. 2007. Selectivity in K+ channels is due to topological control of thepermeant

ions coordination state.Proc. Natl. Acad. Sci. USA 104:926065

9. Bowers KJ, Chow E, Xu H, Dror RO, Eastwood MP, et al. 2006. Scalable algorithms for molecular

dynamics simulations on commodity clusters.Proc. Intl. Conf. for High Performance Computing, Networking

Storage and Analysis(SC06). New York: IEEE

10. Bowers KJ,Dror RO,Shaw DE.2007. Zonal methods forthe parallel executionof range-limited N-body

simulations.J. Comp. Phys.221:30329

11. Buch I, Giorgino T, De Fabritiis G. 2011. Complete reconstruction of an enzyme-inhibitor binding

process by molecular dynamics simulations. Proc. Natl. Acad. Sci. USA 108:101848912. Case DA, Cheatham TE III, Darden T, Gohlke H, Luo R, et al. 2005. The Amber biomolecular

simulation programs.J. Comp. Chem.26:166888

13. Chen J, Brooks CL III, Khandogin J. 2008. Recent advances in implicit solvent-based methods for

biomolecular simulations.Curr. Opin. Struct. Biol.18:14048

14. Chow E, Rendleman CA, Bowers KJ, Dror RO, Hughes DH, et al. 2008. Desmond performance on a

cluster of multicore processors.D. E. Shaw Research Technical Report DESRES/TR2008-01, D. E. Shaw

Research, New York, NY

448 Dror et al.


21/27

15. Denning EJ, Priyakumar DU, Nilsson L, Mackerell AD Jr. 2011. Impact of 2-hydroxyl sampling on the

conformational properties of RNA: update of the CHARMM all-atom additive force field for RNA.J.

Comput. Chem.32:192943

16. Dror RO, Arlow DH, Borhani DW, Jensen M, Piana S, Shaw DE. 2009. Identification of two distinct

inactive conformations of the2-adrenergic receptor reconciles structural and biochemical observations.

Proc. Natl. Acad. Sci. USA106:46899417. Transitions bet

conformational stat

unbiased simulatio

a GPCR suggest a

mechanism for

drug-induced

activation.

17. Dror RO, Arlow DH, Maragakis P, Mildorf TJ, Pan AC, et al. 2011. Activation mechanism of

the 2-adrenergic receptor.Proc. Natl. Acad. Sci. USA 108:1868489

18. Dror RO, Grossman JP, Mackenzie KM, Towles B, Chow E, et al. 2010. Exploiting 162-nanosecondend-to-end communication latency on Anton. Proc. Conf. for High Performance Computing, Networking,

Storage and Analysis(SC10). New York: IEEE

19. Dror RO, Jensen M, Borhani DW, Shaw DE. 2010. Exploring atomic resolution physiology on a

femtosecond to millisecond timescale using molecular dynamics simulations.J. Gen. Physiol. 135:555

6220. Drugs bind

spontaneously in

simulation, revealin

sequence of

intermediate states

route to the

crystallographic bin

pose.

20. Dror RO, Pan AC, Arlow DH, Borhani DW, Maragakis P, et al. 2011. Pathway and mechanism

of drug binding to G proteincoupled receptors.Proc. Natl. Acad. Sci. USA 108:1311823

21. Enkavi G, Tajkhorshid E. 2010. Simulation of spontaneous substrate binding revealing the binding

pathway and mechanism and initial conformational response to GlpT.Biochemistry49:110514

22. Ensign DL, Kasson PM, Pande VS. 2007. Heterogeneity even at the speed limit of folding: large-scale

molecular dynamics study of a fast-folding variant of the villin headpiece. J. Mol. Biol.274:80616

23. Faraldo-G omez J-D, Roux B. 2007. On the importance of a funneled energy landscape for the assembly

and regulation of multidomain Src tyrosine kinase.Proc. Natl. Acad. Sci. USA 104:1364348

24. FineR, Dimmler G, LevinthalC. 1991.FASTRUN:a special purpose,hardwired computerfor molecular

simulation.Proteins11:24253

25. Fitch BG, Rayshubskiy A, Eleftheriou M, Ward TJC, Giampapa M, et al. 2006. Blue Matter: strong

scaling of molecular dynamics on Blue Gene/L. Comp. Sci.3992:84654

26. Freddolino PL, Arkhipov AS, Larson SB, McPherson A, Schulten K. 2006. Molecular dynamics simu-

lations of the complete satellite Tobacco mosaic virus. Structure14:43749

27. Freddolino PL, Liu F, Gruebele M, Schulten K. 2008. Ten-microsecond molecular dynamics simulation

of a fast-folding WW domain.Biophys. J.94:L7577

28. Freddolino PL, Schulten K. 2009. Common structural transitions in explicit-solvent simulations of villin

headpiece folding.Biophys. J.97:233847

29. Friedrichs MS, Eastman P, Vaidyanathan V, Houston M, Legrand S, et al. 2009. Accelerating molecular

dynamic simulation on graphics processing units.J. Comput. Chem.30:8647230. Harvey MJ, Giupponi G, De Fabritiis G. 2009. ACEMD: accelerating biomolecular dynamics in the

microsecond time scale.J. Chem. Theory Comput. 5:163239

31. Hess B, Kutzner C, van der Spoel D, Lindahl D. 2008. GROMACS 4: algorithms for highly efficient,

load-balanced and scalable molecular simulation.J. Chem. Theory Comput. 4:43547

32. Hornak V, Abel R, Okur A, Strockbine B, Roitberg A, Simmerling C. 2006. Comparison of multiple

Amber force fields and development of improved protein backbone parameters.Proteins65:71225

33. Hurst DP, Grossfield A, Lynch DL, Feller S, Romo TD, et al. 2010. A lipid pathway for ligand binding

is necessary for a cannabinoid G protein-coupled receptor. J. Biol. Chem.285:1795464

34. Simulations of a

potassium channel

elucidate the

mechanisms of ionpermeation and vol

gating.

34. Jensen M, Borhani DW, Lindorff-Larsen K, Maragakis P, Jogini V, et al. 2010. Principles of

conduction and hydrophobic gating in K+ channels.Proc. Natl. Acad. Sci. USA 107:583338

35. Jensen M, Dror RO, Xu H, Borhani DW, Arkin IT, et al. 2008. Dynamic control of slow water

transport by aquaporin 0: implications for hydration and junction stability in the eye lens. Proc. Natl.Acad. Sci. USA105:1443035

36. Jiang W, Hardy DJ, Phillips JC, Mackerell AD Jr, Schulten K, Roux B. 2011. High-performance scalable

moleculardynamicssimulations of a polarizable forcefield basedon classical Drude oscillators in NAMD.

J. Phys. Chem. Lett. 2:8792

37. JorgensenWL, Bollini M, ThakurVV, Domaoal RA,Spasov KA,AndersonKS. 2011. Efficientdiscovery

of potent anti-HIV agents targeting the Tyr181Cys variant of HIV reverse transcriptase. J. Am. Chem.

Soc.133:1568696



22/27

38. Jorgensen WL, Buckner JK, Boudon S, Tirado-Rives J. 1988. Efficient computation of absolute free

energies of binding by computer simulations: application to the methane dimer in water. J. Chem. Phys

89:374246

39. Kamerlin SCL, Warshel A. 2011. The empirical valence bond model: theory and applications. Wiley

Interdisciplinary Rev. Comput. Mol. Sci.1:3045

40. Kirkwood JG. 1935. Statistical mechanics of fluid mixtures.J. Chem. Phys.3:30013

41. Kiss G, Rothlisberger D, Baker D, Houk KN. 2010. Evaluation and ranking of enzyme designs. Protein

Sci.19:176073

42. Klauda JB, Venable RM, Freites JA, OConnor JW, Tobias DJ, et al. 2010. Update of the CHARMMall-atom additive force field for lipids: validation on six lipid types. J. Phys. Chem. B144:783043

43. Kryshtafovych A, Fidelis K, Moult J. 2011. CASP9 results compared to those of previous CASP experi-

ments.Prot. Struct. Funct. Bioinform. 79(Suppl. 10):196207

44. Kumar S, RosenbergJM, Bouzida D, Swendsen RH,KollmanPA. 1992. Theweightedhistogram analysis

method for free-energy calculations on biomolecules. I. The method. J. Comput. Chem.13:101121

45. Ladurner AG, Itzhaki LS, Daggett V, Fersht AR. 1998. Synergy between simulation and experiment in

describing the energy landscape of protein folding.Proc. Natl. Acad. Sci. USA 95:847378

46. Lefkowitz RJ, Sun J-P, Shukla AK. 2008. A crystal clear view of the 2-adrenergic receptor. Nat

Biotechnol.26:18991

47. Lei H, Wu C, Liu H, Duan Y. 2007. Folding free-energy landscape of villin headpiece subdomain from

molecular dynamics simulations.Proc. Natl. Acad. Sci. USA 104:492530

48. Lindorff-Larsen K, Best RB, DePristo MA, Dobson CM, Vendruscolo M. 2005. Simultaneous determi-nation of protein structure and dynamics.Nature433:12832

49. Lindorff-LarsenK, MaragakisP, PianaS, EastwoodMP, DrorRO, ShawDE. 2012.Systematic validation

of protein force fields against experimental data.PLoS ONE.7:e32131

50. Simulations of 12

structurally diverse

proteins reproduce

experimentally observed

structures and provide

insight into folding

mechanisms.

50. Lindorff-Larsen K, Piana S, Dror RO, Shaw DE. 2011. How fast-folding proteins fold.Science

334:51720

51. Lindorff-Larsen K, Piana S, Palmo K, Maragakis P, Klepeis JL, et al. 2010. Improved side-chain torsion

potentials for the Amber ff99SB protein force field.Proteins78:195058

52. Lyman E, Higgs C, Kim B, Lupyan D, Shelley JC, et al. A role for a specific cholesterol interaction in

stabilizing the Apo configuration of the human A2Aadenosine receptor.Structure17:166068

53. Mackerell AD Jr, Feig M, Brooks CL III. 2004. Extending the treatment of backbone energetics in

protein force fields: limitations of gas-phase quantum mechanics in reproducing protein conformationa

distributions in molecular dynamics simulations.J. Comput. Chem.25:14001554. Martinez-Mayorga K, Pitman MC, Grossfield A, Feller SE, Brown SA. 2006. Retinal counterion switch

mechanism in vision evaluated by molecular simulations. J. Am. Chem. Soc.128:165023

55. McCammon JA, Gelin BR, Karplus M. 1977. Dynamics of folded proteins.Nature267:58590

56. McDowell SE, Spackova N, Sponer J, Walter NG. 2007. Molecular dynamics simulations of RNA: an

in silico single molecule approach.Biopolymers85:16984

57. Mongan J, Case DA. 2005. Biomolecular simulations at constant pH.Curr. Opin. Struct. Biol.15:15763

58. Monticelli L, Kandasamy SK, Periole X, Larson RG, Tieleman DP, Marrink S-J. 2008. The MARTINI

coarse grained force field: extension to proteins.J. Chem. Theory Comput. 4:81934

59. Morra G, Baragli C, Colombo G. 2010. Selecting sequences that fold into a defined 3D structure: a new

approach for protein design based on molecular dynamics and energetics.Biophys. Chem.146:7684

60. Moukhametzianov R, Warne T, Edwards PC, Serrano-Vega MJ, Leslie AG, et al. 2011. Two distinct

conformations of helix 6 observedin antagonist-bound structures of a1-adrenergic receptor.Proc. NatlAcad. Sci. USA108:822832

61. Nguyen H, Jager M, Kelly J, Gruebele M. 2005. Engineering a -sheet protein towards the folding

speed limit.J. Phys. Chem. B109:1518286

62. NiesenMJ, Bhattacharya S, Vaidehi N. 2011. Theroleof conformational ensemblesin ligandrecognition

in G-protein coupled receptors.J. Am. Chem. Soc.133:13197204

63. Simulations suggest

the physical

mechanisms underlying

ion selectivity in

potassium channels. 63. Noskov SY, Berneche S, Roux B. 2004. Control of ion selectivity in potassium channels by

electrostatic and dynamic properties of carbonyl ligands. Nature431:83034

450 Dror et al.


23/27

64. Perez A, Marchan I, Svozil D, Sponer J, Cheatham TE, et al. 2007. Refinement of the AMBER force

field for nucleic acids improving the description of/ conformers.Biophys. J.92:381729

65. Phillips JC, Braun R, Wang W, Gumbart J, Tajkhorshid E, et al. 2005. Scalable molecular dynamics

with NAMD.J. Comp. Chem.26:1781802

66. Multiple GPUs

used to accelerate t

NAMD parallel MD

software.

66. Phillips JC, Stone JE, Schulten K. 2008. Adapting a message-driven parallel application to GPU-

accelerated clusters.Proc. Int. Conf. High Performance Computing, Networking, Storage and Anal-

ysis(SC08). New York: IEEE

67. Piana S, Lindorff-Larsen K, Shaw DE. 2011. How robust are protein folding simulations with respect

to force field parameterization?Biophys. J.100:L474968. Piana S, Sarkar K, Lindorff-Larsen K, Guo M, Gruebele M, Shaw DE. 2011. Computational design and

experimental testing of the fastest-folding -sheet protein.J. Mol. Biol.405:4348

69. Plimpton S. 1995. Fast parallel algorithms for short-range molecular dynamics.J. Comp. Phys.117:119

70. Ponder JW, Wu C, Ren P, Pande VS, Chodera JD, et al. 2010. Current status of the AMOEBA polar-

izable force field.J. Phys. Chem. B114:254964

71. RasmussenSGF,Choi H-J, Fung JJ,Pardon E, Casarosa P, et al.2011. Structureof a nanobody-stabilized

active state of the 2adrenoceptor.Nature469:17580

72. Romo TD, Grossfield A, Pitman MC. 2010. Concerted interconversion between ionic lock substates of

the 2adrenergic receptor revealed by microsecond timescale molecular dynamics.Biophys. J.98:7684

73. Rosenbaum DM, Zhang C, Lyons JA, Holl R, Aragao D, et al. 2011. Structure and function of an

irreversible agonist-2adrenoceptor complex.Nature469:23640

74. Sanbonmatsu KY, Joseph S, Tung CS. 2005. Simulating movement of tRNA into the ribosome duringdecoding.Proc. Natl. Acad. Sci. USA 102:1585459

75. Schames J, Henchman RH, Siegel JS, Sotriffer CA, Ni H, McCammon JA. 2004. Discovery of a novel

binding trench in HIV integrase. J. Med. Chem.47:187981

76. Seeliger MA, Ranjitkar P, Kasap C, Shan Y, Shaw DE, et al. 2009. Equally potent inhibition of c-Src

and Abl by compounds that recognize inactive kinase conformations.Cancer Res.69:238492

77. Senn HM, Thiel W. 2009. QM/MM methods for biomolecular systems. Angew. Chem. Int. Ed. Engl.

48:119822978. Drug molecule

diffuse from solutio

into their binding s

simulation,

spontaneously adop

the crystallographi

observed conforma

78. Shan Y, Kim ET, Eastwood MP, Dror RO, Seeliger MA, Shaw DE. 2011. How does a drug

molecule find its target binding site? J. Am. Chem. Soc. 133:918183

79. Shan Y, Klepeis JL, Eastwood MP, Dror RO, Shaw DE. 2005. Gaussian split Ewald: a fast Ewald mesh

method for molecular simulation.J. Chem. Phys.122:054101

80. Shan Y, Seeliger MA, Eastwood MP, Frank F, Xu H, et al. 2009. A conserved protonation-dependentswitch controls drug binding in the Abl kinase. Proc. Natl. Acad. Sci. USA 106:13944

81. Shaw DE. 2005. A fast, scalable method for the parallel evaluation of distance-limited pairwise particle

interactions.J. Comp. Chem.26:131828

82. Reports the

performance of An

special-purpose pa

supercomputer, an

completion of the fi

millisecond-long M

simulation.

82. Shaw DE, Dror RO, Salmon JK, Grossman JP, Mackenzie KM, et al. 2009. Millisecond-scale

molecular dynamics simulations on Anton.Proc. Conf. High Performance Computing, Networking,

Storage and Analysis(SC09). New York: ACM

83. MD simulation

long as a millisecon

capture conformati

changes, folding, an

unfolding of severa

proteins.

83. Shaw DE, Maragakis P, Lindorff-Larsen K, Piana S, Dror RO, et al. 2010. Atomic-level charac-

terization of the structural dynamics of proteins.Science330:34146

84. SeibertMM,PatrikssonA,HessB,vanderSpoelD.2005.Reproduciblepolypeptidefoldingandstructure

prediction using molecular dynamics simulations.J. Mol. Biol.354:17383

85. Simmerling S, Strockbine B, Roitberg AE. 2002. All-atom structure prediction and folding simulations

of a stable protein. J. Am. Chem. Soc.124:112585986. Snir M. 2004. A note on N-body computations with cutoffs.Theor. Comput. Sys.37:295318

87. Snow CD,SorinEJ, Rhee YM, Pande VS. 2005. Howwell can simulation predict protein folding kinetics

and thermodynamics?Annu. Rev. Biophys. Biomol. Struct.34:4369

88. Sosnick TR, Barrick D. 2011. The folding of single domain proteinsHave we reached a consensus?

Curr. Opin. Struct. Biol.21:1224

89. Stern HA. 2007. Molecular simulation with variable protonation states at constant pH.J. Chem. Phys.

126:164112



24/27

90. Taiji M, Narumi T, Ohno Y, Futatsugi N, Suenaga A, et al. 2003. Protein explorer: a petaflops special-

purpose computer system for molecular dynamics simulations.Proc. ACM/IEEE Conf. on Supercomputing

(SC03). New York: ACM

91. Tajkhorshid E, Nollert P, Jensen M, Miercke LJW, OConnell J, et al. 2002. Control of the selectivity

of the aquaporin water channel family by global orientational tuning. Science296:52530

92. Toyoda S, Miyagawa H, Kitamura K, Amisaki T, Hashimoto E, et al. 1999. Development of MD engine

high-speed accelerator withparallel processordesign for moleculardynamics simulations.J. Comp. Chem

20:18599

93. van der Kamp MW, Schaeffer RD, Jonsson AL, Scouras AD, Simms AM, et al. 2010. Dynameomics: acomprehensive database of protein dynamics.Structure18:42335

94. van Duin ACT, Dasgupta S, Lorant F, Goddard WA III. 2001. ReaxFF: a reactive force field for hydro-

carbons.J. Phys. Chem. A105:9396

95. Vanni S, Neri M, Tavernelli I, Rothlisberger U. 2009. Observation of ioniclockformation in molecular

dynamics simulations of wild-type 1 and 2 adrenergic receptors.Biochemistry48:478997

96. Vanommeslaeghe K, Hatcher E, Acharya C, Kundu S, Zhong S, et al. 2010. CHARMM general force

field (CGenFF): a force field for drug-like molecules compatible with the CHARMM all-atom additive

biological force fields.J. Comp. Chem.31:67190

97. Voelz VA, Bowman GR, Beauchamp K, Pande VS. 2011. Molecular simulation of ab initio protein

folding for a millisecond folder NTL9(139).J. Am. Chem. Soc.132:152628

98. Wang J, Wolf RM, Caldwell JW, Kollman PA, Case DA. 2004. Development and testing of a general

Amber force field.J. Comp. Chem.25:11577499. Woo H-J, Roux B. 2005. Calculation of absolute proteinligand binding free energy from computer

simulations.Proc. Natl. Acad. Sci. USA 102:682530

100. Yang S, Banavali NK, Roux B. 2009. Mapping the conformational transition in Src activation by cu-

mulating the information from multiple molecular dynamics trajectories. Proc. Natl. Acad. Sci. USA

106:377681

101. Yang W, Pan Y, Zheng F, Cho H, Tai HH, Zhan CG. 2009. Free-energy perturbation simulation on

transition states and redesign of butyrylcholinesterase.Biophys. J.96:193138

102. Yildirim I, Stern HA, Tubbs JD,Kennedy SD,Turner DH.2011. BenchmarkingAMBERforce fields for

RNA: Comparisons to NMR spectra for single-stranded r(GACC) are improved by revised torsions

J. Phys. Chem. B115:926170

103. Zeevaart JG, Wang L, Thakur VV, Leung CS, Tirado-Rives J, et al. 2008. Optimization of azoles

as anti-human immunodeficiency virus agents guided by free-energy calculations. J. Am. Chem. Soc130:949299

104. Zgarbova M, Otyepka M, Sponer J, Mladek A, Banas P, et al. 2011. Refinement of the Cornell et al

nucleic acids force field based on reference quantum chemical calculations of glycosidic torsion profiles

J. Chem. Theory Comput. 7:2886902

105. Zimmerberg J, Bezanilla F, Parsegian VA. 1990. Solute inaccessible aqueous volume changes during

opening of the potassium channel of the squid giant axon. Biophys. J.57:104964

106. Zuckerman DM. 2011. Equilibrium sampling in biomolecular simulations.Annu. Rev. Biophys.40:4162

107. Zwanzig RW. 1954. High-temperature equation of state by a perturbation method. I. Nonpolar gases

J. Chem. Phys.22:142026

452 Dror et al.


25/27

Annual Review o

Biophysics

Volume 41, 2012Contents

How Should We Think About the Ribosome?

Peter B. Moore 1

Mechanisms of Sec61/SecY-Mediated Protein Translocation

Across MembranesEunyong Park and Tom A. Rapoport 21

Racemic Protein CrystallographyTodd O. Yeates and Stephen B.H. Kent 41

Disulfide Bonding in Protein Biophysics

Deborah Fass 63

Prokaryotic Diacylglycerol Kinase and Undecaprenol Kinase

Wade D. van Horn and Charles R. Sanders 81

Allostery and the Monod-Wyman-Changeux Model After 50 YearsJean-Pierre Changeux 103

Protein Structure in Membrane DomainsArianna Rath and Charles M. Deber 135

Bacterial Mechanosensitive ChannelsMscS: Evolutions Solution

to Creating Sensitivity in FunctionJames H. Naismith and Ian R. Booth 157

Cooperativity in Cellular Biochemical Processes: Noise-Enhanced

Sensitivity, Fluctuating Enzyme, Bistability with Nonlinear

Feedback, and Other Mechanisms for Sigmoidal

Responses

Hong Qian 179

Network-Based Models as Tools Hinting at Nonevident

Protein FunctionalityCanan Atilgan, Osman Burak Okan, and Ali Rana Atilgan 205

Filamins in Mechanosensing and SignalingZiba Razinia, Toni Makela, Jari Ylanne, and David A. Calderwood 227

vii


26/27

ATP Util

Documents

Biomolecular Simulation: A Computational Microscope for Molecular Biology